E-TRENDS: Enhanced LSTM Trend Forecasting for Equities

Battle Count: 119

Highly relevant to quantitative trading. The paper directly addresses systematic trend-following strategies, provides a formal theoretical framework (bias-variance theorem for differenced forecasting), benchmarks against standard ML models used in practice (OLS, Ridge, Lasso, LightGBM), and evaluates economic performance via portfolio simulations with transaction costs. The Sharpe-ratio loss function directly optimizes for trading performance. However, the absolute Sharpe ratios remain near zero or negative, and the improvement over baseline, while statistically significant, may not translate to strong absolute alpha generation.

Moderate complexity. The LSTM architecture is relatively simple (2 layers, 64 units, dropout). Feature engineering involves standard technical indicators (SMA, RSI, MACD, t-statistics). The main complexity lies in: (1) proper chronological data splitting to avoid lookahead bias, (2) hyperparameter grid search across multiple dimensions, (3) implementing the custom Sharpe-ratio loss function, (4) portfolio-level simulation with transaction costs, and (5) stationarity testing and preprocessing. The Colab notebook reduces implementation barriers significantly.

Full analysis encapsulated in an interactive Google Colab notebook covering data preprocessing, model training, evaluation, and visualization. All Python dependencies installable via single cell. Users can modify hyperparameters interactively. Data sourced from Yahoo Finance (publicly available). However, no formal GitHub repository is provided, and the exact random seeds and full training logs are not explicitly stated.